增加交叉验证和超参数调优

前文中，只是给了基础模型：

PyTorch 实现 CIFAR-10 图像分类：从数据预处理到模型训练与评估-CSDN博客

今天我们增加交叉验证和超参数调优，

先看运行结果：
===== 在测试集上评估最终模型 =====
最终模型在测试集上的准确率：60.14%
最优模型已保存为 'cifar10_best_model.pth'（超参数：{'batch_size': 32, 'epochs': 5, 'lr': 0.01, 'momentum': 0.85}）

Process finished with exit code 0
比基础模型准确率高了一点，

完整代码如下：

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader, Subset
import matplotlib.pyplot as plt
import numpy as np
import torchvision
from sklearn.model_selection import KFold, ParameterGrid  # 用于交叉验证和超参数网格搜索# --------------------------
# 1. 数据准备（与原代码一致，但后续会在训练集内部做交叉验证）
# --------------------------
# 数据预处理：标准化（与原代码相同）
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])# 数据集路径（请替换为你的实际路径）
data_path = r'D:\workspace_py\deeplean\data'# 加载完整训练集和测试集（测试集始终不变，用于最终评估）
full_trainset = datasets.CIFAR10(root=data_path, train=True, download=False, transform=transform)
testset = datasets.CIFAR10(root=data_path, train=False, download=False, transform=transform)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')# --------------------------
# 2. 定义CNN模型（与原代码一致）
# --------------------------
class Net(nn.Module):def __init__(self):super(Net, self).__init__()self.conv1 = nn.Conv2d(3, 6, 5)self.pool = nn.MaxPool2d(2, 2)self.conv2 = nn.Conv2d(6, 16, 5)self.fc1 = nn.Linear(16 * 5 * 5, 120)self.fc2 = nn.Linear(120, 84)self.fc3 = nn.Linear(84, 10)def forward(self, x):x = self.pool(torch.relu(self.conv1(x)))x = self.pool(torch.relu(self.conv2(x)))x = x.view(-1, 16 * 5 * 5)x = torch.relu(self.fc1(x))x = torch.relu(self.fc2(x))x = self.fc3(x)return x# --------------------------
# 3. 交叉验证函数（核心新增）
# --------------------------
def cross_validate(model, train_dataset, k_folds=5, epochs=5, lr=0.001, batch_size=32, momentum=0.9):"""5折交叉验证：将训练集分成5份，每次用4份训练，1份验证，返回平均准确率"""kfold = KFold(n_splits=k_folds, shuffle=True, random_state=42)  # 固定随机种子，结果可复现fold_results = []  # 存储每折的验证准确率for fold, (train_ids, val_ids) in enumerate(kfold.split(train_dataset)):print(f'\n===== 第 {fold + 1}/{k_folds} 折交叉验证 =====')# 1. 划分当前折的训练集和验证集train_subset = Subset(train_dataset, train_ids)  # 本次训练用的数据val_subset = Subset(train_dataset, val_ids)  # 本次验证用的数据# 2. 创建数据加载器train_loader = DataLoader(train_subset, batch_size=batch_size, shuffle=True)val_loader = DataLoader(val_subset, batch_size=batch_size, shuffle=False)# 3. 初始化模型和优化器（每折都重新训练新模型，避免干扰）model_instance = Net()  # 重新实例化模型criterion = nn.CrossEntropyLoss()optimizer = optim.SGD(model_instance.parameters(), lr=lr, momentum=momentum)# 4. 训练当前折的模型for epoch in range(epochs):model_instance.train()  # 训练模式running_loss = 0.0for i, data in enumerate(train_loader, 0):inputs, labels = dataoptimizer.zero_grad()outputs = model_instance(inputs)loss = criterion(outputs, labels)loss.backward()optimizer.step()running_loss += loss.item()# 每200步打印一次损失（简化输出）if i % 200 == 199:print(f'折 {fold + 1}，轮次 {epoch + 1}，第 {i + 1} 步：平均损失 {running_loss / 200:.3f}')running_loss = 0.0# 5. 在验证集上评估当前折的模型model_instance.eval()  # 验证模式correct = 0total = 0with torch.no_grad():for data in val_loader:images, labels = dataoutputs = model_instance(images)_, predicted = torch.max(outputs.data, 1)total += labels.size(0)correct += (predicted == labels).sum().item()val_acc = 100 * correct / totalprint(f'第 {fold + 1} 折验证准确率：{val_acc:.2f}%')fold_results.append(val_acc)# 计算所有折的平均准确率（该超参数组合的最终得分）avg_acc = sum(fold_results) / len(fold_results)print(f'\n===== 该超参数组合的平均验证准确率：{avg_acc:.2f}% =====')return avg_acc# --------------------------
# 4. 超参数调优（核心新增）
# --------------------------
def hyperparameter_tuning(train_dataset):"""超参数网格搜索：尝试不同的超参数组合，用交叉验证选最优"""# 定义要测试的超参数组合（可根据需要增减）param_grid = {'lr': [0.001, 0.01],  # 学习率：尝试两个值'batch_size': [32, 64],  # 批大小：尝试两个值'momentum': [0.9, 0.85],  # 动量：尝试两个值'epochs': [5]  # 训练轮次（固定为5，减少计算量）}best_acc = 0.0best_params = None  # 存储最优超参数# 遍历所有超参数组合（共 2×2×2=8 种组合）for params in ParameterGrid(param_grid):print(f'\n---------- 测试超参数组合：{params} ----------')# 用交叉验证评估当前组合的性能current_acc = cross_validate(model=Net(),train_dataset=train_dataset,k_folds=5,epochs=params['epochs'],lr=params['lr'],batch_size=params['batch_size'],momentum=params['momentum'])# 记录最优组合if current_acc > best_acc:best_acc = current_accbest_params = paramsprint(f'★ 发现更优组合！当前最优准确率：{best_acc:.2f}%')print(f'\n===== 超参数调优完成 =====')print(f'最优超参数：{best_params}')print(f'最优平均验证准确率：{best_acc:.2f}%')return best_params# --------------------------
# 5. 主函数：执行超参数调优 + 最终训练 + 测试集评估
# --------------------------
if __name__ == '__main__':# 步骤1：超参数调优（用交叉验证选最优参数）print('===== 开始超参数调优（这一步比较慢，需要耐心等待）=====')best_params = hyperparameter_tuning(full_trainset)# 步骤2：用最优超参数在完整训练集上训练最终模型print('\n===== 用最优超参数训练最终模型 =====')final_model = Net()criterion = nn.CrossEntropyLoss()optimizer = optim.SGD(final_model.parameters(),lr=best_params['lr'],momentum=best_params['momentum'])train_loader = DataLoader(full_trainset,batch_size=best_params['batch_size'],shuffle=True)# 训练最终模型（轮次与调优时一致）for epoch in range(best_params['epochs']):final_model.train()running_loss = 0.0for i, data in enumerate(train_loader, 0):inputs, labels = dataoptimizer.zero_grad()outputs = final_model(inputs)loss = criterion(outputs, labels)loss.backward()optimizer.step()running_loss += loss.item()if i % 200 == 199:print(f'最终模型训练 - 轮次 {epoch + 1}，第 {i + 1} 步：平均损失 {running_loss / 200:.3f}')running_loss = 0.0# 步骤3：在测试集上评估最终模型（用从未见过的测试数据）print('\n===== 在测试集上评估最终模型 =====')final_model.eval()test_loader = DataLoader(testset, batch_size=32, shuffle=False)correct = 0total = 0with torch.no_grad():for data in test_loader:images, labels = dataoutputs = final_model(images)_, predicted = torch.max(outputs.data, 1)total += labels.size(0)correct += (predicted == labels).sum().item()test_acc = 100 * correct / totalprint(f'最终模型在测试集上的准确率：{test_acc:.2f}%')# 步骤4：保存最优模型torch.save(final_model.state_dict(), 'cifar10_best_model.pth')print(f"最优模型已保存为 'cifar10_best_model.pth'（超参数：{best_params}）")